KataGo Distributed Training and new networks

And · Post by **And** » Sat Nov 26, 2022 8:13 pm

KataGo 18b, playouts 100 - CS Zero 9d H4

And · Post by **And** » Sun Nov 27, 2022 8:16 am

KataGo 18b, playouts 100 - CS Zero 9d H5
after 347 moves CS Zero: winrate 64.3, B+32

PC_Screen · Post by **PC_Screen** » Sun Nov 27, 2022 6:23 pm

New KataGo 18b out, supposed to as strong as 60b at equal visits while also being faster than 40b, which means it should be stronger than both in an actual game.

https://cdn.discordapp.com/attachments/ ... 21b.bin.gz

Also test game: KataGo 1.11 18b (W) vs KataGo 1.3.5 40b (B) with 2H and 0.5 Komi, Japanese Rules. The time setting is 10s/move for both which gave them around 1-2k visits per move on average.

The new 18b KataGo managed a convincing win against a 2 year old KataGo despite the 2 handicap stones.

And · Post by **And** » Sun Nov 27, 2022 6:27 pm

"UEC finished this weekend...KataGo placed 2nd in the final (and first in the preliminaries)." https://discord.com/channels/4170221623 ... 4930181303

And · Post by **And** » Mon Nov 28, 2022 12:47 pm

KataGo (18b-uec, playouts 1) - CS Zero 9d

And · Post by **And** » Mon Nov 28, 2022 9:37 pm

100 games 19x19 KataGo 18b-uec (playouts 1) - 60b-s665 (playouts 1):
18b (B) - 60b (W) 21:29
18b (W) - 60b (B) 26:24

And · Post by **And** » Tue Nov 29, 2022 12:05 am

100 games 19x19 KataGo 18b-uec (playouts 2) - 60b-s665 (playouts 1):
18b (B) - 60b (W) 32:18
18b (W) - 60b (B) 35:15

And · Post by **And** » Tue Nov 29, 2022 6:32 am

10 games KataGo 18b-uec (playouts 10) - 60b-s665 (playouts 5):
18b (W) - 60b (B) 8:2

And · Post by **And** » Tue Nov 29, 2022 10:52 am

10 games KataGo 18b-uec (playouts 10) - 60b-s665 (playouts 5):
18b (B) - 60b (W) 9:1

And · Post by **And** » Tue Nov 29, 2022 11:28 am

in the position from here viewtopic.php?p=244851#p244851, I turned on the analysis mode, the network 18b-uec. the move h1 was immediately highlighted. after ~200 visits, katago showed white winning, but then decided that there no winning and the winrate continued to fall

EDIT after 10K visits - 39.4

Click Here To Show Diagram Code: [go]$$Wc White to play $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . O . . X . . . . . . . . . | $$ | . X O O . . . X X O X X X . X X X . . | $$ | . . X , O . O X O O . O . . . O X O . | $$ | . X . X X X O O X O . . . . . . . . . | $$ | . . O . . . O . X O . . X . . . O . . | $$ | . . . . . O . . X . O O X . . O . . . | $$ | . . X . . . O X X X O X X . X O . . . | $$ | . . . . . . O O X O O X . O . X O . . | $$ | . . . , . . O O O X X O . . . X O . . | $$ | . . . . . . X O O X . O . . . X . . . | $$ | . . O . . X . X X X O . . . . . X . . | $$ | . . . . . X O . . X . . . . . . . . . | $$ | . . . X . X O O O X . . . . . . . . . | $$ | . . X O X X X O X O . O . . . . O . . | $$ | . . O O O . X O X O . . . O . O . O . | $$ | . . . O . O O X X O . . X X X X O X . | $$ | . . . . . . X . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

PC_Screen · Post by **PC_Screen** » Tue Nov 29, 2022 8:45 pm

Result of a 100 games match between the latest 18b vs the strongest 40b (S1184), 1s/move (averaging 200-500 visits/move), Chinese Rules and komi is set to 7.0 which allows for draws. A draw is counted as +0.5 points for both players:
18b vs 40b:

65 : 35

From 18b's perspective:
Wins: 58
Draws: 14
Losses: 28

Elo Difference: +126.51 ± 73.59

And · Post by **And** » Wed Nov 30, 2022 5:00 am

10 games KataGo 18b-uec (playouts 20) - 60b-s665 (playouts 10):
18b (B) - 60b (W) 9:1

And · Post by **And** » Fri Dec 02, 2022 10:21 am

10 games KataGo 18b-uec (playouts 20) - 60b-s665 (playouts 10):
18b (W) - 60b (B) 10:0

PC_Screen · Post by **PC_Screen** » Sat Dec 03, 2022 9:25 am

Thorough test done by @Medwin on the Go AI Discord, using 400 visits/move with a total of 2800 games played.

: image.png (350.38 KiB) Viewed 23534 times

As expected, the new b18-uec is slightly stronger than 60b at equal visits (much stronger with equal time) and is expected to grow even stronger as its training run progresses. When KataGo 1.12 releases 18b is expected to replace 60b as the main network for training games.

PC_Screen · Post by **PC_Screen** » Tue Dec 06, 2022 2:26 pm

Result of a 100 games match between the latest 18b vs the strongest 20b, 400 visits/move, Chinese Rules and komi is set to 7.0 which allows for draws. A draw is counted as +0.5 points for both players:
18b vs 20b:

92.5 : 7.5

From 18b's perspective:
Wins: 92
Draws: 1
Losses: 7

Elo Difference: +447.48 ± 132.32

Also did the same against Leela Zero, 1s/move, Tromp-Taylor Rules, 7.5 Komi, and the result was:

18b vs LZ 40B:

100:0

Elo can't be calculated from this since 18b didn't lose any games. Most estimates put Leela ~800 elo or so behind KataGo

Example Games:

18b vs 20b

18b vs LZ 40b:

Life In 19x19

KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks

Re: KataGo Distributed Training and new networks